ISOM 260 · Suffolk University

Agent Patterns &
Production Intelligence

ReAct reasoning, multi-agent architectures, and the patterns that separate demos from production systems

ReAct Multi-Agent MCP Business Automation
Use arrow keys to navigate
Part 1 · 5:00–5:12 PM

The Pattern Gap

Bridge from Session 5's NanoClaw demo — the patterns behind production agents.

The Competence Ladder

S4
Build agents with tool use
Function calling, structured outputs, first agent loop
S5
Connect to real APIs + memory
Live Wikipedia, web fetch, conversation persistence
S6
ReAct + Multi-Agent + MCP
Production patterns, team architectures, universal tool standard
NanoClaw Callback
When NanoClaw researched for us live — it didn't call one tool and stop. It searched, evaluated, searched again, then answered. That's a PATTERN.
ReAct
Think before acting. Observe results. Adapt and iterate.
Multi-Agent
Team of specialists collaborating on complex tasks.
MCP
Universal tool standard. Build once, every agent uses it.
Part 2 · 5:12–5:40 PM

ReAct Deep Dive

The pattern that powers every serious AI agent: Reason + Act. Think, then do, then evaluate, then decide what's next.

THINK
Reason about what to do next
→
ACT
Call a tool with purpose
→
OBSERVE
Evaluate the result
→
DECIDE
Repeat or conclude

The One-Line Difference

Session 5 — System Prompt
"You are a helpful research assistant. You have access to tools for searching Wikipedia, fetching web pages, and doing calculations."
vs
Session 6 — ReAct Prompt
"You are a research analyst. For every question, follow this loop: THINK: What do I need to find out? ACT: Call the best tool for this step. OBSERVE: What did I learn? Is it enough? REPEAT or CONCLUDE with a synthesis. Always show your reasoning."
Key Insight
The agent loop is identical to Session 5. The magic is in the prompt, not the code.
Three Levels

Three Levels of Intelligence

Same question, three approaches. Same tools for Levels 1 & 2 — watch the escalation.

L0 No Tools
Generates text from training data. Looks confident — no way to verify.
# Just asks the LLM directly "AI affects 40% of jobs..." # Real? 🤷 No sources.
❌ Looks good. Can't verify.
L1 Session 5 Agent
Has tools. Calls one, dumps the result. Real data, but no evaluation.
wikipedia_search("AI") # Gets first result "Here's what Wikipedia says..." # Grabbed first match. Done.
⚠️ Real data, first result only.
L2 ReAct Agent
Same tools. Thinks, searches, evaluates, searches more, synthesizes.
THINK: Need employment data wikipedia_search("Tech_unemp") OBSERVE: Good, need specifics wikipedia_search("AI_jobs") OBSERVE: Better. Verify... fetch_web_page("...") THINK: Enough. Synthesize.
✅ Planned. Evaluated. Synthesized.
Same tools. Same loop. One prompt paragraph changed everything.
Hands-On · 5:40–6:25 PM

Build a ReAct Research Agent

45 minutes in Google Colab. Same tools, new system prompt, dramatically different behavior.

Open Colab Now
API key ready. We're building a research agent that thinks before it acts.
Open in Google Colab

Workshop Phases

Failure Modes

When Agents Fail

Production agents need to handle the messy real world. Here are four scenarios that break naive agents.

Fake Company
Search for "Nexorium Technologies" — a company that doesn't exist. Does the agent admit "no data found" or hallucinate a plausible-sounding answer?
Vague Opinion Question
Ask "Is AI good?" — a question tools can't answer. Does the agent search endlessly or recognize it's a judgment call?
Sources Disagree
When Wikipedia says one thing and a web source says another. Does the agent pick one, acknowledge the conflict, or dig deeper?
Infinite Loop Risk
The agent keeps searching for "perfect" data that doesn't exist. Does it know when to stop? Does it have a max-iteration guard?
What to watch for: Graceful degradation, honest uncertainty, and knowing when to stop. The best agents say "I don't know" when they don't know.
Part 5 · 6:35–7:00 PM

Multi-Agent Architectures

One agent is useful. A team of agents is powerful. Three patterns for organizing agent teams.

Pipeline — Assembly Line

Researcher
web_search, read_pdf
→
Writer
draft_doc, cite
→
Editor
check_facts, score
Each agent does its part and passes to the next. Like a factory assembly line.

Debate — Adversarial Review

Proposer
generates answer
⇄
Critic
finds flaws
→
Synthesizer
best of both
Arguments and counterarguments. Adversarial pressure improves quality.

Orchestrator — Manager + Workers

Manager Agent
dispatches · monitors · combines
↙ ↓ ↘
Data Worker
sql_query, chart
Research Worker
web_search, read
Writing Worker
draft, format
Hub-and-spoke model. Just like companies have departments — the manager delegates.
Design exercise: Pick a business process, sketch a 3-agent pipeline on paper. Which pattern fits? Where does a human need to approve?
MCP

MCP — The USB-C of AI

The Model Context Protocol: an open standard for connecting AI agents to tools and data sources. Build once, every agent uses it.

Before MCP
N × M
Every agent needs a custom integration for every tool. 5 agents × 10 tools = 50 custom connectors.
After MCP
N + M
Each agent speaks MCP. Each tool speaks MCP. 5 agents + 10 tools = 15 connections.
Build once, every agent uses it
1,000+ MCP servers already exist. Slack, GitHub, Google Drive, databases, and more. Browse them at modelcontextprotocol.io

Popular MCP Servers

Slack GitHub Google Drive PostgreSQL Notion Gmail Filesystem Brave Search
NanoClaw's ipc-mcp.ts: Our live demo agent connects to tools via MCP — the same standard that powers Claude Code, Cursor, and hundreds of other AI tools.
Part 6 · 7:00–7:18 PM

Business Automation & Agent ROI

Where do agents create real business value? A framework for deciding what to automate, augment, or leave to humans.

The Automation Matrix

← Low Judgment · High Judgment →
AUTOMATE
High Volume + Low Judgment
Invoice processing, data entry
AUGMENT
High Volume + High Judgment
Sales outreach, support triage
DON'T BOTHER
Low Volume + Low Judgment
Rare admin tasks
HUMAN
Low Volume + High Judgment
Strategy, crisis decisions
↑ High Volume · Low Volume ↓

ROI Formula

Hours Saved × Hourly Cost × Error Reduction − Agent Cost = Monthly Value
The Trust Equation
Agent autonomy scales inversely with stakes.
Low stakes → let the agent run autonomously.
High stakes → human in the loop, every time.
Quick discussion: What task from your job or internship would you automate first? Where on the matrix does it land?
Part 7 · 7:18–7:30 PM

The Road Ahead

Where we've been, where we're going, and why the agentic arc is the future of this course.

S1-2
AI
Fundamentals
S3-4
Prompts &
Products
S5
Real-World
Agents
S6
Agent
Patterns
S7
RAG
S8
Multi-Agent
Build
S9
Industry
Agents
S10
Guard-
rails
S11
Eval
S12+
Final
Project

Key Takeaways

ReAct = Think before you act. The best agents don't just execute — they reason about what to do, observe results, and adapt. This pattern separates toy demos from production systems.
Agent teams beat solo agents. Just like human organizations, specialized agents collaborating on complex tasks produce better results than one agent trying to do everything.
MCP is the USB-C of AI. The Model Context Protocol means you build a tool integration once and every agent can use it. This standard is reshaping how agents connect to the world.
Automate the routine, augment the judgment. The Automation Matrix helps you decide what to fully automate, what to augment with AI, and what must stay human.
The code changed one paragraph. The thinking changed everything. Your Session 5 agent — same tools, same loop — became a research powerhouse with one system prompt change. That's the ReAct pattern.
"The code changed one paragraph.
The thinking changed everything."
Homework

Homework & Resources

Extend your agent, critique its research, design a multi-agent system, and explore MCP.

Extend
Add a 4th tool to your research agent (e.g., competitor_lookup, financial_data, or industry_news). Test it with 3 different business questions and document the reasoning traces.
Analyze
Pick one research output from your agent. Write a 1-page critique: Were the sources credible? Did the reasoning have gaps? What would a human analyst do differently?
Design
Write a 1-page multi-agent system proposal for a business process you care about. Include: 3+ agent roles, tools for each, handoff logic, human review gates, and estimated ROI.
Explore
Browse modelcontextprotocol.io and find 3 MCP servers relevant to your final project idea. Write one paragraph each on how they would help your agent.
Read
Read Anthropic's "Building Effective Agents" guide. Write 3 key takeaways and identify one scenario where you should NOT use an agent.

Resources

Advanced Agent Patterns
Visual explainer: ReAct, CoT, multi-agent, and MCP.
ReAct Agent Walkthrough
Interactive animated ReAct trace with real data.
Anthropic Agent Docs
Official guide to building agentic systems with Claude.
Model Context Protocol
Browse 1,000+ MCP servers. The USB-C of AI.
Next Session
Session 7: Knowledge Agents — RAG
Your agent is smart, but it doesn't know YOUR business. Add company knowledge with Retrieval-Augmented Generation.